Analysing accuracy, balancing bias: Can ChatGPT be used to ease the care documentation burden?

GoLTC Data Science Group

May 2023

GOLTC Data Science

  • GOLTC Introduction Adelina Comas-Herrera
  • Data Science Interest Group: Structure and webinars
    • Steering Group: Sam Rickman (CPEC at LSE), Jiyoun Song (UPenn), Tommy Henderson-Reay (NHS England), Sarthak Saluja (CPEC at LSE)
  • Racial disparities and harmful hallucinations in automated speech recognition Allison Koenecke, Cornell Department of Information Science
  • Adapted large language models can outperform medical experts in clinical text summarization Dave Van Veen, Stanford Center for Artificial Intelligence

Documentation burden

  • An assessment or intervention takes place with a person using care services.
  • Later the worker types up a summary on a case recording system.
  • 6 - 20 hours per worker per week spent on writing documentation:
    • Reduced time to spend care and assessment.
    • High levels of burnout across professions.

Case study: Magic Notes from Beam

  1. Conversation is recorded using a worker’s phone.
  2. This audio file is transcribed into text using an AI speech-to-text model.
  3. This transcript is summarised into a record of the meeting using an AI summarisation model, such as GPT4.

Today’s webinar

  1. Disparities in automated speech recognition (Allison Koenecke, Cornell Department of Information Science)
  2. Adapted large language models can outperform medical experts in clinical text summarization (Dave Van Veen, Stanford Center for Artificial Intelligence)

Each presentations will be around 20 minutes, followed by 10 minutes of questions.